Abstract
Since the description ischuria renalis by William Heberden (1), AKI has remained a prominent complication of critical illness. Beyond KRT, treatment has been limited by the capacity to phenotype this condition. Here, we chronicle the evolution of attempts to classify AKI, including the adoption of consensus definitions, the expansion of diagnosis and prognosis with novel biomarkers, and emerging tools such as artificial intelligence (AI).
The Need for a Consensus Definition
Classic textbooks characterize AKI as a rapid decline in GFR and retention of nitrogenous waste products (2). Although clinically accurate, lack of specificity has spawned >30 definitions, hindering comparisons between settings (3,4). To address these limitations, the Risk, Injury, Failure, Loss, and End Stage Renal Disease (RIFLE) scheme (Figure 1), initially proposed by the Acute Dialysis Quality Initiative (ADQI) in 2004, provided a standardized framework to classify AKI. The least severe category “risk” was defined as a 50% increase in serum creatinine or >25% decrease in eGFR presumed to have occurred within 7 days, with increases in severity labeled “injury,” “failure,” “loss,” and “ESRD” with parallel criteria for oliguria (5). The subsequent Acute Kidney Injury Network (AKIN) Classification System in 2005 eliminated RIFLE’s Loss and ESRD stages and eGFR criteria, and added KRT to stage 3 (6). Given studies showing mortality associated with small increases in creatinine (7), AKIN included a 0.3 mg/dl creatinine increase over 48 hours to improve sensitivity. The most recent AKI classification scheme, introduced by the Kidney Disease Improving Global Outcomes (KDIGO) workgroup in 2012 (8), combined elements of both: stage 1 AKI criteria could be met by an increase in creatinine of either 0.3 mg/dl within 48 hours or a 50% increase within 7 days with comparisons showing equivalent or superior sensitivity for AKI detection and similar prognostic performance for AKI staging and outcomes (9⇓⇓–12). Validation studies have also shown that the urine output criteria alone (13,14) and in combination with serum creatinine is associated with mortality in critically ill patients (15). Direct comparisons of urine output and creatinine criteria suggest that creatinine may be the more predictive of the two, with studies in cardiac surgery and cardiac intensive care unit (ICU) patients demonstrating higher risk for adverse outcomes (prolonged length of stay, KRT, and mortality) with AKI by creatinine criteria compared with equivalent stage AKI by urine output (16,17). AKI by both creatinine and urine output criteria had even higher risk of KRT and prolonged length of stay compared with AKI by creatinine alone (16). Despite the prognostic value of urine output, logistic and interpretative challenges have hindered its greater application and generalizability. The hourly capture of this information is challenging in patients without a catheter and outside the ICU and can be confounded by diuretic use and solute loading. Additionally, the optimal threshold urine output is unclear, with one study of patients undergoing abdominal surgery finding a risk of AKI associated with an intraoperative urine output <0.3 ml/kg per hour, but not with urine output <0.5 ml/kg per hour (18). Although routine surveillance may be challenging, short-term monitoring of urine output with provocative testing has shown promise as a prognostic marker in milder AKI. The “furosemide stress test,” assessed by urine flow rate after a standardized dose of furosemide, has been shown to predict progression to higher stages of AKI (19,20).
RIFLE, AKIN, and KDIGO systems for AKI classification. Urine output and serum creatinine criteria for the RIFLE, AKIN, and KDIGO classification systems for AKI. AKIN, Acute Kidney Injury Network; Cr, creatinine; KDIGO, Kidney Disease Improving Global Outcomes; RIFLE, Risk, Assessment, Failure, Loss, and End Stage Renal Disease; SCr, serum creatinine.
What Have the Consensus Definitions Taught Us about AKI in the ICU
The application of these definitions uncovered the unappreciated burden of kidney dysfunction among the critically ill. Before RIFLE, reported incidences of AKI ranged from 1% to 25%, although these studies primarily focused on more severe AKI (3,4,21). Studies utilizing the RIFLE, AKIN, or KDIGO criteria in ICU populations reported higher incidences, with one- to two-thirds of patients afflicted (11,22⇓⇓⇓⇓–27). Consensus definitions have also elucidated the prognosis of AKI beyond earlier studies, which demonstrated in-hospital mortality rates of up to 40%–60% in severe cases (3,4). By leveraging a consistent framework, studies have shown that mild AKI is associated with mortality (7,28) and a graded association between increasing AKI severity and death (Table 1) (22,23,27,29⇓⇓⇓⇓–34). The former demonstrated that even in the absence of overt failure, lesser degrees of kidney dysfunction were clinically relevant, facilitating widespread adoption of the term “acute kidney injury.” Although not limited to critical illness, studies using these definitions have also shown associations between AKI and long-term sequelae including kidney disease progression (35,36), cardiovascular outcomes (37⇓–39), and frailty (40) (Table 1).
Studies of AKI outcomes
These consensus definitions have also provided a framework to study the dynamic trajectory of AKI during critical illness and the importance of recovery. In one study, patients who recovered from AKI within 7 days and did not experience a relapse had 1-year survival >90% (41), whereas those who had a relapse of AKI after early recovery had a five-fold higher risk of death at 1 year. No recovery from AKI had the worst prognosis with approximately 40% survival at 1 year (41). The extent to which these associations are explained by a lack of kidney recovery itself versus the more severe underlying illness prompting nonrecovery is not clear; however, these findings parallel studies of AKI survivors in large health care databases demonstrating an association between the degree of recovery and the risk for mortality and CKD over longer durations (35,42).
Use in Clinical Trials and Practice
AKI management recommendations on the basis of stage have been proposed in the KDIGO practice guidelines (8) and the National Kidney Foundation AKI Core Curriculum (43); however, benefit in routine practice remains to be demonstrated. More recently, the KDIGO criteria have been applied to standardize and improve the reproducibility of clinical trials and enrich for patients more likely to benefit from interventions. Recent studies comparing early versus late initiation of dialysis used these criteria to identify more advanced stages of AKI to enrich for participants more likely to require dialysis (44⇓⇓–47). These criteria have also been increasingly applied as outcomes in studies of balanced crystalloid use (48,49) and trials evaluating “bundle” interventions for AKI prevention (50,51). As familiarity improves, standard use of these criteria may also help audiences interpret and compare the effect of interventions.
Limitations of Consensus Definitions
Despite improving understanding, conceptual and logistic limitations remain. AKI requires observed changes in creatinine, but the “baseline” creatinine to anchor that definition is often missing. For patients admitted with elevated creatinine, it can be unclear whether the dysfunction is chronic or acute. Various strategies are used to impute surrogate values, each with strengths and limitations (Table 2), and this variety has hindered effective comparisons when applying consensus definitions (5,52⇓–54). Although modifications to introduce specificity and “reduce the need for a baseline” (e.g., use of a 48-hour window) have been proposed, the potential for underestimating AKI incidence or severity remains (Table 2). Further, creatinine and urine output can be confounded by changes in the volume of distribution, diuretic use, and altered creatinine production, which can threaten specificity and project an overly optimistic picture of recovery (55⇓⇓–58). One study of patients with ICU stays >5 days showed persistent decreases in serum creatinine in both AKI and non-AKI patients (59). These limitations may be most apparent when using the most sensitive definitions of AKI (e.g., stage 1). Notably, the original RIFLE criteria acknowledged that these criteria may be overly sensitive in attempting to capture the larger population at “risk” for parenchymal injury; however, labeling mild changes as “injury” in subsequent iterations may have unintendedly masked this original sentiment. Many of these limitations, and some emerging concepts, were noted in the 2019 KDIGO controversies conference of AKI, including the need for subsequent definitions to include distinctions for AKI persistence, transience, and relapsing or recovered AKI, and the potential incorporation of kidney injury biomarkers (60).
Perhaps the most insidious limitation of the consensus definitions is the temptation to mistake the precision they add as a replacement that addresses the larger phenotyping barriers still limiting progress in developing novel therapeutics in AKI. From the application of existing care strategies to better aligning novel interventions to target populations, the need for a pragmatic and accurate way to classify the heterogeneity and pathophysiology underlying AKI remains a critical challenge.
Acute Kidney Disease
It has also been recognized that AKI and CKD are not discrete entities, but exist in a continuum, with short reversible changes identified as AKI and persistent or irreversible changes in kidney function identified as de novo or progressive CKD. Although current AKI and CKD definitions are useful constructs for studying epidemiology, their limitations include failure to recognize smaller or more subacute changes in kidney function that are associated with clinically relevant outcomes or account for the prognostic importance of kidney function trajectory after an AKI event. In practice, the absence of an inclusive definition may delay recognition and treatment of acute kidney disorders not meeting AKI or CKD criteria but with common risk factors and outcomes.
To address these gaps, the term “acute kidney disease” (AKD) was first proposed as part of the 2012 KDIGO AKI definition to better identify all kidney injury (biomarker positivity, including proteinuria) or functional derangements lasting <90 days. In this paradigm, AKI—by virtue of placing a greater emphasis on events occurring within a shorter time frame (48 hours to 7 days)—is a subset of AKD. Given the heterogenous nature of AKD, an initial classification system and management recommendations were not included. Non-AKI AKD could have several clinical phenotypes in both inpatient and outpatient settings, including newly identified biomarker positivity (proteinuria or novel biomarkers) in the absence of changes in kidney function, newly identified abnormalities in kidney function in the absence of a reference value, and subacute changes in kidney function (not meeting temporal AKI criteria).
AKD has not been systematically studied, due in part to the recognition that the term could represent many different phenotypes. James et al. described the incidence and prognostic importance of AKD in a Canadian population-based cohort study that included both hospitalized and nonhospitalized patients and found that AKD without AKI was three times more prevalent than AKI (3.8% versus 1.5%) and associated with a higher risk of CKD, kidney failure, and mortality compared with patients with no kidney disease (61). Similar findings have been shown using UK and Danish cohorts (62,63).
In 2017 the ADQI proposed an alternative definition of AKD that included a classification system and enhanced integration within the current AKI framework (64). In this schema, AKI was identified by acute changes in kidney function occurring within a 7-day time frame, with persistent changes in kidney function lasting beyond 7 days labeled as AKD (Figure 2). AKD could then be classified according to AKI stage to better classify the extent of kidney recovery and its prognostic effect after AKI. Changes persisting beyond 90 days would be reclassified as CKD. Although potentially useful in the setting of critical illness, where kidney function is frequently measured and etiology and epidemiology are well studied, frequent assessment of kidney function is lacking in most studies, with the lack of electronic health record (EHR) interoperability in the United States representing one challenge for follow-up. The framework requires further validation and does not address subacute kidney injury (non-AKI AKD), which clearly has prognostic importance. The KDIGO AKD definition was recently revised to reflect differences between AKD without AKI and AKD with AKI, with staging on the basis of eGFR and albuminuria added in those with AKD and not AKI (65). As our understanding of kidney injury evolves, our definitions of AKI and AKD will likely evolve in parallel.
AKI, acute kidney disease, and chronic kidney disease continuum. Framework proposed by the Acute Dialysis Quality Initiative (ADQI) workgroup for AKI, acute kidney disease (AKD), and CKD. Duration of time after the injury determines the category of disease (AKI, AKD, or CKD). Stages of AKD map to the corresponding AKI stages. Stage 0 AKD indicates partial recovery from AKI. Stage 0A includes patients who have completely recovered from AKI with no residual evidence of injury but retain the risk of long-term events. Stage 0B includes patients whose serum creatinine has returned to baseline but who have evidence of ongoing kidney damage, injury, or loss of kidney functional reserve. Stage 0C includes patients with serum creatinine levels above baseline but within 1.5 times baseline. Figure and caption from the consensus report of the ADQI 16 workgroup by Chawla et al. (64). KDIGO, Kidney Disease Improving Global Outcomes; SCr, serum creatinine.
Novel Biomarkers of AKI
The above attempts have collectively underscored sensitivity and specificity limitations of serum creatinine and urine output, a point recently illustrated by a biopsy series of kidney damage where one-third of patients did not meet KDIGO AKI criteria (66). These limitations have generated interest in the development and validation of novel biomarkers that better reflect parenchymal injury and provide prognostic information. Novel markers have generally been characterized as indicating tubular “damage” or “stress,” with the former including, but not limited to, neutrophil gelatinase-associated lipocalin (NGAL), kidney injury molecule–1 (KIM-1), IL-18, and L-type fatty acid binding protein (L-FABP), and the latter including tissue-inhibitor of metalloproteinases 2 (TIMP-2) and IGF-binding protein 7 (IGFBP7). Other candidate biomarkers, such as C-C motif chemokine ligand 14 (CCL14), dickkopf-related protein 3 (DKK3), and chitinase 3–like–1 gene product (YKL-40), have emerged as potentially promising predictors of postoperative AKI (67), AKI persistence (68), and AKI progression and hospital mortality (69). These have led to novel concepts such as “subclinical injury” and characterizing AKI by the presence of functional and/or structural damage (Figure 3) (70). The clinical importance of subclinical injury was demonstrated in a study of the French and European Outcome Registry in ICUs (FROG-ICU) and Adrenomedullin and Outcome in Severe Sepsis and Septic Shock-1 (AdrenOSS-1) cohorts of critically ill patients, in which subclinical AKI (defined by elevated plasma proenkephalin A 119–159 levels) occurred in 6.1%–6.7% of patients and was associated with a higher risk of death at 28 days compared with those without subclinical AKI (hazard ratio [HR] 2.4; 95% confidence interval [95% CI], 1.5 to 3.7, for FROG-ICU cohort; HR 2.5; 95% CI, 1.1 to 5.9, for AdrenOSS-1 cohort) (71). Other areas of investigation have included risk stratification, early detection of AKI, assessing prognosis and recovery, enriching clinical trials, detecting nephrotoxic signals, and measuring responses to interventions.
AKI categorization using novel biomarkers. At the tenth consensus conference of the Acute Dialysis Quality Initiative (ADQI) workgroup, an early framework for the clinical use of novel biomarkers in combination with functional biomarkers of AKI (i.e., serum creatinine and urine output) was developed, wherein functional change and evidence of structural damage (manifested by elevations in damage biomarkers) together provided more detailed categorization of AKI. This framework was recently updated during the 23rd ADQI meeting. In the revised system, each KDIGO stage of AKI is subcategorized by the presence or absence of damage biomarkers. For KDIGO stage 1 AKI (shown in the figure), these include stage 1S (i.e., “subclinical” AKI), stage 1A (functional change without elevated damage biomarkers), and stage 1B (functional change and elevated damage biomarkers). Stages 2–3 AKI are defined by parallel criteria. These changes aim to improve the sensitivity for AKI detection and to discriminate underlying etiology and assess severity, although this staging system for AKI remains to be validated (91). Notably, subclinical AKI has been shown to be associated with adverse outcomes, including mortality and need for KRT (112,113). Red arrows denote progression; blue arrows denote resolution. Figure adapted from Koyner et al. (114) and Ostermann et al. (91). KDIGO, Kidney Disease Improving Global Outcomes.
A focus of early inquiry examined the potential of these markers to “predict” or detect early AKI, with studies demonstrating mixed performance (72⇓⇓⇓⇓⇓⇓–79). Two recently discovered markers, TIMP-2 and IGFBP7, are expressed in kidney tubular cells during physiologic stress or injury. Their combined use has been demonstrated to predict the development of stage 2–3 AKI within 12 hours in critically ill patients with areas under the curve between 0.79 and 0.82, generally outperforming earlier markers of tubular injury (80,81). Although the Food and Drug Administration allowed marketing as a complementary tool for risk assessment in adults with recent cardiovascular and respiratory failure (82), studies attempting to better define the clinical interpretation and actionability of TIMP-2*IGFBP7 and other markers remain. Some limitations include challenges with comparison with an imperfect creatinine standard (83⇓–85) and difficulty pinpointing the exact timing of injury in the critically ill (86,87), which may help explain the superior negative predictive values relative to the more modest positive predictive values generally observed (88). Despite some uncertainty regarding their role for early diagnosis, several biomarkers have been shown to provide additional prognostic information (Table 3).
Studies of novel biomarkers
Notably, because kidney biopsies are rarely performed in the ICU, few studies have compared the performance of these markers against a histologic standard. One study examined the performance of L-FABP, IL-18, and KIM-1 for diagnosing acute tubular injury (ATI) on biopsy (89). Compared with serum creatinine, NGAL levels were higher in mild and severe ATI compared with no ATI and exhibited higher discrimination for severe ATI (0.67; 95% CI, 0.60 to 0.74) compared with creatinine alone (0.58; 95% CI, 0.49 to 0.67) (89). Although not in a critically ill population, these findings suggest that the specificity for some markers for detecting histologic ATI may be modest. Efforts to continue validating these markers and find clinicopathologic correlates are ongoing. One example is the Kidney Precision Medicine Project, an ambitious program whose goal is to prospectively enroll patients with AKI and CKD to define molecular pathways for specific disease subphenotypes through the collection of biospecimens and the development of a kidney tissue atlas (90).
Given these collective characteristics, the ADQI recently suggested that damage markers be combined with conventional functional markers to improve diagnostic accuracy and assess severity, including a modification to KDIGO staging that adds substages on the basis of biomarker levels (91). The overall strength of the recommendation was a B (conditional), indicating that further research is needed to improve confidence. Nevertheless, many of these markers have been leveraged in recent clinical trials to enrich for populations at higher risk for clinical outcomes (Table 3) (45,50,51). In the Biomarker Guided Implementation of the KDIGO Guidelines to Reduce the Occurrence of AKI in Patients After Cardiac Surgery (PrevAKI) study, an AKI care bundle (optimization of volume status and hemodynamics, avoidance of nephrotoxins, and prevention of hyperglycemia) was tested in patients undergoing cardiac surgery among those with an elevated level of TIMP-2*IGFBP7 (≥0.3) after cardiopulmonary bypass and demonstrated reduced incidence of AKI (Table 3) (51). Similarly, the Biomarker-guided Intervention to Prevent Acute Kidney Injury After Major Surgery (BigpAK) study tested an AKI bundle in patients undergoing major abdominal surgery and used TIMP-2*IGFBP7 levels to identify ICU patients at high risk of AKI. Although no global differences in AKI stages between arms were observed, a subgroup with elevated biomarker levels showed a reduction in AKI incidence (27% versus 48%; P=0.03) (50). These studies indicate that biomarker-guided interventions may be useful in identifying patients who may respond to clinical interventions. Whether these performance characteristics can extend to routine practice for triage purposes, or prompt more aggressive diagnostic evaluation or resuscitation, is under investigation (92,93).
Another interest has been in determining whether biomarkers can discriminate between known phenotypes of AKI and help discover new phenotypes. Few studies have compared expression patterns of biomarkers to differentiate causes of AKI. In one study of 218 patients who underwent kidney biopsies for evaluation of AKD, high levels of TNF-α and IL-19 were strongly associated with acute interstitial nephritis, enhancing the area under receiver operating characteristic curve of prebiopsy clinical impression (94). Recently, a substudy of the Vasopressin and Septic Shock Trial in patients with AKI measured angiopoietin-1 and -2, soluble tumor necrosis factors receptor 1, and IL-18 and used latent class analysis to identify two subphenotypes of AKI, one of which had a higher risk of nonrecovery and 28-day mortality, suggesting that similar approaches could be used to identify molecularly distinct AKI subphenotypes with differential responses to therapy (95).
In summary, novel biomarkers for AKI remain an evolving area of investigation. Despite ongoing molecular and clinical validation of their strength as specific indicators of tubular injury or stress, findings to date have led to early applications suggesting potential use cases in clinical trials, in phenotyping, and in guiding current clinical management.
Artificial Intelligence/Informatics
The simplicity of laboratory-based consensus definitions of AKI makes them an ideal target for interventions using clinical decision support (CDS). Leveraging EHRs using bioinformatics and AI has potential uses in the context of AKI, including AKI alerting, predictive analytics, AKI phenotyping, and risk-based management. The feasibility of CDS/AI to reduce the rate, duration, and intensity of AKI has been demonstrated in a number of clinical settings (96,97). Not all studies using CDS/AI have demonstrated improved outcomes (98) however, perhaps reflecting differences in implementation and clinical heterogeneity.
Predicting AKI using CDS paired with specific interventions may be more promising. The Nephrotoxic Injury Negated by Just in time Action (NINJA) alert program is a notable example, where Goldstein and colleagues demonstrated a sustained reduction in nephrotoxin-mediated AKI in hospitalized pediatric patients by identifying patients at highest risk through nephrotoxic drug exposure and deprescription (99). The NINJA alert program has been validated in multiple centers (100).
AI has been used to predict AKI in numerous care settings (101⇓–103). Koyner et al. utilized EHR data to develop a prediction tool that predicts stage 2 AKI a median of 41 hours before a rise in serum creatinine and the need for dialysis within 2 days. The tool performed well across numerous adult care settings (102). Siminov et al. implemented an AKI prediction tool across several institutions and were able to predict AKI 24 hours before a rise in creatinine and the need for dialysis and mortality (103). Other groups have used similar techniques to predict AKI in specific clinical settings, including adults with severe burns (104), and acute pancreatitis (105), postoperative (106), and cardiac patients (107,108). Although such prediction tools have potential to mitigate ongoing kidney injury, all require external validation and their implementation has not been shown to improve patient-centered outcomes. Evidence-based interventions supported by these risk-prediction tools need to be tested before widespread implementation.
AI may have a further role in phenotyping AKI trajectory and risk assessment after an AKI event, including prediction of risk of rehospitalization, recurrent AKI, heart failure hospitalization, and other comorbidities. Semler et al. demonstrated the feasibility of using AI to predict major adverse kidney events by 30 days (MAKE 30), which may allow delivery of risk-stratified post-AKI follow-up care in the future (109).
There are many opportunities to leverage the power of AI to recognize and phenotype AKI and to improve the care and follow-up of AKI patients. The AI working group of the AKI!Now initiative of the American Society of Nephrology represents the growing momentum in this field and ongoing collaborative efforts to use data science and quality initiatives to improve early recognition and treatment of AKI and to reduce the disease burden on patients and health systems (110,111).
Conclusion
Critically ill patients are subject to diverse and severe forms of AKI. Although the past three decades have brought significant advances in AKI epidemiology and prognosis, the development of treatment for AKI beyond supportive (i.e., dialytic) therapy has been stalled by the limited phenotyping of this heterogeneous condition. There is reason for optimism, however, because continued multifaceted approaches to advance AKI phenotyping may eventually allow investigators to better pair potential therapies with underlying pathophysiology and identify new targets.
Disclosures
N. Pannu reports employment with Alberta Health Services; consultancy agreements with GE; serving on the board of directors of the Kidney Foundation of Canada, Northern Alberta branch; and funding from Amgen for a quality improvement initiative in ESKD. N. Pannu also reports honoraria from Astellas in 2011; World Congress of Nephrology (WCN)—covered travel expenses as speaker $1000.00 in 2013; International Society of Nephrology (ISN)—covered travel expenses of $700.00 for Mexican National Nephrology Meeting March 2015; Acute Dialysis Quality Initiative (ADQI)—travel expenses for November 2015, June 2017 meeting; WCN—speaker, covered expenses 2017, 2019, 2020; and ISN—speaker, covered expenses 2017, 2018, 2019, ADQI AKI meeting 2020 $1000.00 USD honorarium, and AKI Continuous Renal Replacement Therapy meeting 2021 (provided microphone and lamp for speakers). E.D. Siew reports employment with Nashville Veterans Affairs, consultancy agreements with Akebia Therapeutics on 4/2019, honorarium for an invited educational talk on AKI epidemiology at the DaVita Annual Physician Leadership Conference 2/2019, serving as an Associate Editor of CJASN, and royalties as an author for UpToDate. The remaining author has nothing to disclose.
Funding
B.C. Birkelo is supported by National Institute of Diabetes and Digestive and Kidney Diseases grant T32DK007569-32. N. Pannu is supported by funding from the Canadian Institute for Health Research. E.D. Siew is supported by the Vanderbilt O’Brien Kidney Center grant P30-DK114809 Clinical and Translational Research Core.
Acknowledgments
Because Dr. Edward D. Siew is an Associate Editor of CJASN, he was not involved in the peer review process for this manuscript. Another editor oversaw the peer review and decision-making process for this manuscript.
Author Contributions
Bethany C. Birkelo was responsible for conceptualization, data curation, methodology, and visualization; wrote the original draft; and reviewed and edited the manuscript. Neesh Pannu was responsible for conceptualization, data curation, investigation, methodology, and visualization; wrote the original draft; and reviewed and edited the manuscript. Edward D. Siew was responsible for conceptualization, investigation, methodology, project administration, supervision, and visualization; wrote the original draft; and reviewed and edited the manuscript.
Footnotes
Published online ahead of print. Publication date available at www.cjasn.org.
- Copyright © 2022 by the American Society of Nephrology